Mahesh ramachandran

Fast sfm

We develop a fast, convergent, scalable and parallelizable algorithm for structure from motion assuming the availability of additional measurements about the camera motion. Along with the image sequence obtained from a moving camera, we assume that we have measurements of the gravity vector (vertical direction) and camera height from the ground.

 

Using these measurements, we simplify the SFM equations and rewrite them in a bilinear form in the unknown variables. We propose an iterative strategy composed of  (1) depth iterations, (2) motion iterations and (3) side-information refinement iterations to solve for the structure and motion variables.

 

We analyze the computational complexity of the algorithm and conclude that it is of the order of

O(no. of cameras x no. of points). The algorithm can be implemented on a parallel architectures enabling the solution to SFM of very long sequences. In this respect, the algorithm is scalable.

 

We show the algorithm’s application to large datasets of the size that is common of Google’s StreetView datasets. We show results on some sequences in the Google Research dataset that was made available to the research community.

 

Please find illustrative video sequences here: (descriptions may be found in the publications).